Banglaasr

🚀 Bangla ASR Model

A Bangla Automatic Speech Recognition (ASR) model fine - tuned on the Bangla Mozilla Common Voice Dataset.

🚀 Quick Start

The Bangla ASR model is fine - tuned from the Whisper model using the Bangla Mozilla Common Voice Dataset. It was trained on approximately 400 hours of data, with 40k samples for training and 7k for validation. After 12000 training steps, it achieved a Word Error Rate (WER) of 4.58%.

💻 Usage Examples

Basic Usage

import os
import librosa
import torch
import torchaudio
import numpy as np

from transformers import WhisperTokenizer
from transformers import WhisperProcessor
from transformers import WhisperFeatureExtractor
from transformers import WhisperForConditionalGeneration

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

mp3_path = "https://huggingface.co/bangla-speech-processing/BanglaASR/resolve/main/mp3/common_voice_bn_31515636.mp3"

model_path = "bangla-speech-processing/BanglaASR"


feature_extractor = WhisperFeatureExtractor.from_pretrained(model_path)
tokenizer = WhisperTokenizer.from_pretrained(model_path)
processor = WhisperProcessor.from_pretrained(model_path)
model = WhisperForConditionalGeneration.from_pretrained(model_path).to(device)


speech_array, sampling_rate = torchaudio.load(mp3_path, format="mp3")
speech_array = speech_array[0].numpy()
speech_array = librosa.resample(np.asarray(speech_array), orig_sr=sampling_rate, target_sr=16000)
input_features = feature_extractor(speech_array, sampling_rate=16000, return_tensors="pt").input_features

# batch = processor.feature_extractor.pad(input_features, return_tensors="pt")
predicted_ids = model.generate(inputs=input_features.to(device))[0]


transcription = processor.decode(predicted_ids, skip_special_tokens=True)

print(transcription)

📚 Documentation

Dataset

The model uses the Mozilla Common Voice Dataset, which includes around 400 hours of data for both training (40k samples) and validation (7k samples) in MP3 format. For more information about the dataset, please click here.

Training Model Information

Evaluation

The Word Error Rate (WER) of the model is 4.58%. For more information, please check the github.

📄 License

This project is licensed under the MIT license.

📖 Citation

@misc{BanglaASR ,
  title={Transformer Based Whisper Bangla ASR Model},
  author={Md Saiful Islam},
  howpublished={},
  year={2023}
}